Automatic Tissue Classification in Multispectral Mris via an Unsupervised Model

In this paper we propose a novel hybrid multispectral MR images segmentation framework which combines Markov Random Field (MRF), stochastic relaxation (SR) scheme and an improved Genetic Algorithm (IGA). MRF models have by now been firmly established as a robust tool for de-noising and segmentation of medical images. Characterizing MRFs by their local properties (Markovianity), they are well suited to depict the spatial contextual constrains between voxels. Markovianity allows to achieve the global objective by concentrating on labeling of each site consider to its local neighbors. To determine these characteristics practically, Hammersley-Clifford theorem reduces the problem to the minimization of a non convex energy function. In order to optimize the solution, stochastic relaxation (SR) method has been employed widely. SR is very powerful for searching local regions of the solution space exhaustively via stochastic hill climbing and prevents from exploring the same diversity of solutions additionally has the solution refining capability, but imposes heavy overhead on the algorithm. In comparison, genetic algorithm (GA) has a good capability of global researching and converges quickly to a near-global optimum but is weak in hill climbing. By combining SR and GA for optimization, this paper puts forward a new 3D-MRF model for de-noising and segmentation of brain multispectral MR images. This conjunction allows us to keep the benefits of both SR and GA algorithms while simultaneously addressing their individual drawbacks. Ultimately, the performance of the novel method is investigated on the BrainWeb dataset. Experimental results demonstrate that the proposed method outperforms the traditional 2D-MRF in convergence speed and quality of the solution noticeably which is valuable in processing environment.